Research Article |
|
Corresponding author: Chelsea Gray ( graychelsea13@gmail.com ) Academic editor: Daniel Ayllón
© 2026 Chelsea Gray, Dale S. Rothman, Erin Peters-Burton, Cynthia Smith, E.C.M Parsons, Alexis Garretson.
This is an open access article distributed under the terms of the Creative Commons Attribution License (CC BY 4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Citation:
Gray C, Rothman DS, Peters-Burton E, Smith C, Parsons ECM, Garretson A (2026) An individual-based model of basking sharks in Ireland. Individual-based Ecology 2: e176474. https://doi.org/10.3897/ibe.2.176474
|
Basking sharks (Cetorhinus maximus) are largely solitary, following highly individualized migration routes, yet they exhibit poorly understood aggregation behaviors, ranging from pairs to groups of up to 1,000 individuals. These surface aggregations put sharks at risk of boat strikes and are thought to be driven in part by localized zooplankton availability, though observational evidence suggests that food alone may not fully explain this behavior. Previous modeling approaches have successfully identified long-range basking shark migration trends but have not accurately reproduced behavior at finer spatial scales, and commonly used chlorophyll a proxies for zooplankton have produced inconsistent results when compared with sightings data.
This study presents the first individual-based model (IBM) of basking shark aggregation behavior, simulating shark movement around the Inishowen Peninsula, Ireland, from 1982 to 2018. Rather than relying on a phytoplankton proxy, zooplankton was simulated directly to capture fine-scale prey patchiness. Four behavioral submodels were tested and compared with three decades of sightings data from the Irish Basking Shark Group and Irish Whale and Dolphin Group: sharks seeking food only, sharks seeking other sharks only, sharks seeking both food and other sharks, and a random-movement control.
The combined food-and-social submodel produced the most realistic results, with lower error and greater qualitative correspondence to observed sightings data than the food-only, social-only, or random submodels, which rarely generated aggregations larger than two individuals. These findings indicate that basking shark aggregations emerge from a combination of environmental and social drivers rather than food availability or chance alone. This model demonstrates the viability of IBMs for simulating basking shark behavior, offers a flexible framework that can incorporate improved environmental and behavioral data as they become available, and provides a foundation for future research into the drivers of basking shark aggregation.
Basking sharks, Individual-based model, marine conservation, shark behavior
Basking sharks (Cetorhinus maximus) are the second-largest extant fish in the ocean, the only species in the family Cetorhinidae, and are considered endangered worldwide (
In 2022, due partially to the significance of these aggregations, basking sharks were listed under Section 23 of Ireland’s Wildlife Act of 1976, forbidding disturbance of basking shark feeding and reproductive habits, such as these aggregations. Implementation of strong conservation policies requires an improved understanding of basking shark social behavior, which is critical for prioritizing areas of habitat for basking shark social, reproductive, or feeding potential. Currently, there is a lack of conclusive data surrounding basking shark social and migratory behavior, limiting the potential of temporal or area-based management measures. It is difficult to directly observe these aggregations or collect environmental data in situ, as doing so relies on finding an elusive species. Tagging data, while valuable, are prohibitively expensive and only provide data on a small subset of any population.
Here, the first individual-based model (IBM) of basking shark aggregation behavior is presented. The model’s capacity to reproduce observed aggregation patterns is demonstrated by simulating the period from 1982 to 2018 and comparing four behavioral submodels with three decades of sightings reports from the Irish Basking Shark Group (IBSG) and Irish Whale and Dolphin Group (IWDG).
While IBMs have been used for ecological research, they have never been applied to basking sharks (
Basking sharks are uniquely suited to an IBM because, although they are largely solitary and follow highly individualized migration routes, they exhibit poorly understood and understudied aggregation behaviors (
The IBSG and IWDG have collected publicly reported basking shark sightings for five decades (
The model presented here sought to simulate basking shark movement around Inishowen to determine whether IBMs can produce realistic behavioral results, in particular whether a more natural preyscape (simulated through zooplankton patchiness) can improve the realism and effectiveness of modeling shark behavior. The Inishowen Peninsula was selected due to its proximity to an established marine protected area and the availability of sightings data and to limit the simulated area for simplicity. The results of this model were compared with the IBSG/IWDG sightings dataset.
The IBM simulated daily basking shark movements around Inishowen under three behavioral conditions: 1) sharks only seek food, 2) sharks only seek other sharks, and 3) sharks seek both food and other sharks, to assess whether food availability, social attraction, or a combination of these were sufficient to reproduce the aggregations sighted around the Inishowen Peninsula. It is currently unknown where basking sharks mate or give birth, with only one recorded capture of a pregnant female in the North Atlantic in 1943 (
The model was created in NetLogo (version 6.2.2,
Basking shark sighting reports from the IBSG/IWDG were used for calibration and validation, as they constitute the only comprehensive dataset on basking shark population and distribution in Ireland (
The primary objective of this research was to determine whether IBMs are a workable method for modeling basking shark behavior, demonstrating the viability of such a tool for future research and experimentation. This paper serves as a model description and proof-of-concept approach. To demonstrate the utility of the model, three hypotheses of behavior (sharks seek other sharks, sharks only seek food, and sharks seek a combination of food and other sharks) and a random control were tested using different submodels.
Model development and testing proceeded in four phases: (1) exploratory testing to establish plausible parameter ranges from the literature and small ad hoc trials; (2) parameter selection, in which candidate configurations were compared with sightings data using the mean error (ME), root mean square error (RMSE), and mean absolute error (MAE); (3) confirmatory model runs at full scale (200 sharks, 50 trials) across all four submodels; and (4) sensitivity analysis, robustness analysis, and a parametric stress test on the selected configuration. The IBM is available on GitHub (https://github.com/graychelsea13-byte/Basking_Shark_IBM/).
Each time step in the model represents 24 h. The modeled area ranges from 56°N, -8°W to 55°N, -6.5°W, an area of 10,545 km2, with patches of 1 km × 1 km × 10 m (depth). Patches of 1 km × 1 km were selected to keep patch sizes small but manageable and to account for shark movements within 24 h (
Each simulation depicts the period from 1 April–31 October
The agents in the model are individual basking sharks. The model is run with 200 sharks because, while the largest sighting in Inishowen was 60 sharks, the largest sighting in Ireland was 160 sharks. The sharks are not distinguishable by size, age, or sex due to a lack of available data on population demographics. Sharks do not give birth or die in the model. The patches within the model can be land or water. The water patches are each assigned a unique number of zooplankton. In each 24 h cycle, the total amount of zooplankton in the model is calculated from the CPR data.
Zooplankton data came from a formal data request to the Sir Alister Hardy Foundation for Ocean Science (SAHFOS), which conducts the CPR Survey (
| Parameter | Explanation | Setting |
|---|---|---|
| Threshold_ZP | Minimum amount of zooplankton (Cal and Other_ZP combined) required for a shark to stay in or move to a patch. Count of individual zooplankton. | 0–1000000000000 |
| No_Eat_Min | Number of days a shark must encounter a patch with less than the Threshold_ZP before leaving the model. | 0–100 |
| Sense-Distance | How far a shark can see, smell, or hear (in km). This is doubled when sharks are locating other sharks. | 0–100 |
| Swim-Speed | The distance a shark can swim (in km). | 0–100 |
| Return-Season | Number of days it will take a shark to return after it has left in response to reaching the No_Eat_Min. | 0–100 |
| Cal_% | Percentage of patches with Calanus copepods. | 0–100 |
| Other_ZP_% | Percentage of patches with Pseudocalanus spp. and Centropages typicus. | 0–100 |
| Friend_Min | Number of other sharks a patch must have to attract a shark. | 0–100 |
At each time step, the average abundance of zooplankton across all water patches for the entire model environment is assumed to be given by the linearly interpolated CPR data. Since the CPR samples 3 m3 of water, the values are divided by three to obtain an average value per 1 m3. To distribute the zooplankton across the area, it is first assumed that only a percentage of patches contain each category of zooplankton on a given day; these percentages can be altered with a slider using the parameters Cal_% and Other_ZP_% (see Table
The model assumes annual fidelity for all simulated sharks (200 out of an estimated 10,000 worldwide (
Within the model, the senses of sight, smell, and hearing are collectively called “Sense-Distance.” This is not a fixed parameter but can be altered before the start of the model run, as the mechanism by which basking sharks sense zooplankton is still unclear
Sharks retain a memory of high-zooplankton patches (defined as those that contain three times the threshold level of zooplankton, given by the parameter “Threshold_ZP”) to simulate intermittent site fidelity (demonstrated by field observations), which has been attributed to resource availability (
During each daily time step, each basking shark first decides whether to move to a new patch. This decision is based on two factors: (1) the amount of zooplankton and (2) the number of sharks within its current patch. Sharks determine whether there is sufficient zooplankton to meet the threshold level of zooplankton (Threshold_ZP, which can be altered via a slider) for the total number of sharks in the patch (Fig.
Shark decision pathway for Food, Social, and Food/Social submodels. Shark decision pathway under different submodels. Each day, sharks complete this decision tree. Note that if a patch that meets the condition is identified, under all versions, sharks take the following action: if it is within swimming distance, move to it; if it is beyond swimming distance, swim toward it. The Random control is not included in this figure, as sharks simply select a random patch to move to each day and do not follow any decision pathway related to either food availability or the location of other sharks.
Where and why a shark will move next depend, in part, on the submodel chosen (see Fig.
Assuming a target patch is ultimately found, sharks then determine whether the target patch is within the distance they can swim in one time step (given by the parameter Swim-Speed), which may be smaller than Sense-Distance. If the target patch is within the distance set by Swim-Speed, the shark will move to that patch in a single time step. If it is farther away, the shark will move the maximum distance allowed by the Swim-Speed parameter toward the target patch. Sharks do not automatically continue toward the selected patch after 24 h, as it is assumed that sharks alter their behavior based on new information and always preferentially select closer patches over those that are farther away. If the sharks cannot find a target patch, they will swim in a random direction, following the same procedure to avoid land, for the distance given by Swim-Speed. This overall process is repeated at each time step, even if a shark moved only partway toward its target patch in the previous time step.
At each time step, sharks determine whether there was sufficient zooplankton (both Cal and Other_ZP) to support the shark (i.e., meet the minimum level set by Threshold_ZP). If not, the shark records this as a day on which it did not ‘eat’.
At each time step, the following data are collected: (1) the total number of basking sharks within the model area; (2) for patches containing two or more basking sharks, the number of sharks, the amount of zooplankton in the patch, and the location of the patch; and (3) Total Aggregations, i.e., the number of patches with two or more basking sharks. The model also includes randomized sampling of patches (referred to as Pseudo-Sighting Reports), intended to simulate the IBSG/IWDG public sightings reports by hikers and boaters who frequent the area in and around the Inishowen Peninsula (
Shark decision pathway under different submodels. Each day, sharks complete this decision tree. Note that if a patch that meets the condition is identified, under all versions, sharks take the following action: if it is within swimming distance, move to it; if it is beyond swimming distance, swim toward it. The Random control is not included in this figure, as sharks simply select a random patch to move to each day and do not follow any decision pathway related to either food availability or the location of other sharks.
While all non-behavioral elements of the model remain the same, the model consists of four alternative submodels of shark behavior related to their decisions about target patches once they have decided to move from their current patch (Fig.
Only IBSG/IWDG sightings from 1982–2018, from April through October, were included in the data analysis, with the exception that sightings in the last week of March were included in the April count to increase the dataset available for data analysis.
To compare the model output data with the real-world sightings data, the average number of sharks per aggregation for each month (hereafter referred to as monthly aggregation size) was determined for the Total Aggregations, Pseudo-Sighting Reports, and IBSG/IWDG data. The average monthly number of aggregations was also calculated for all datasets.
All datasets were normalized using minimum–maximum normalization, and the mean error (ME), mean absolute error (MAE), and root mean square error (RMSE) were then calculated between the Total Aggregations and IBSG/IWDG monthly averages and between the Pseudo-Sighting Reports and IBSG/IWDG monthly averages. RMSE, ME, and MAE were used because they are statistical measures used for model forecasting of non-normal and inconsistently patterned datasets (
The normalized average monthly aggregation sizes were also graphed as time series and box plots. Visual comparisons were then made between model outputs and the IBSG/IWDG sightings data. Each model run was compared with two subsets of the IBSG/IWDG sightings data: the full Irish dataset (referred to as “Ireland”) and the subset of sightings reported from within the simulated model area (referred to as “Inishowen”).
Initial exploratory testing to determine appropriate model parameter settings was largely qualitative, with initial parameter values drawn from the existing literature for Threshold_ZP (
Small experimental trials were run under different setting combinations and compared with a predefined definition of realism derived from the IBSG/IWDG data. The majority of basking shark aggregations in Ireland range from 2–10 sharks, with aggregations of 15–40 sharks occurring approximately 15% of the time and a maximum aggregation size of at least 60 but no greater than 160. A range of semi-realistic settings was identified by comparing the aggregation size distribution within the model with the IBSG/IWDG data, an approach consistent with pattern-oriented modeling (
The majority of these experimental trials resulted in unrealistic frequencies of aggregation sizes, with aggregations often not exceeding 20 sharks or dramatically exceeding 160 sharks. Experimentation continued in an ad hoc manner until model results with relatively realistic maximum aggregation sizes and frequencies of aggregation sizes were achieved. Approximately 2,000 model tests were run during this exploratory process.
The results of the exploratory tests that produced realistic aggregation size distributions and realistic maximum aggregation sizes were then compared with the IBSG/IWDG data using RMSE, ME, and MAE analyses. Those tests with the lowest RMSE results when compared with the IBSG/IWDG data
Once settings that produced strong RMSE results were identified, more detailed preliminary testing was conducted to identify the settings that produced the most realistic results for all submodels. Initial calibration testing showed high consistency across trials for the Random and Social submodels but low similarity with the IBSG/IWDG data. This is likely because the Random and Social submodels rarely produced aggregations of more than two or three sharks under any combination of settings. Thus, preliminary tests were conducted using only the Food and Food/Social submodels to save computational power and time. To further reduce computational needs, parameter selection tests were conducted with a maximum of 100 sharks in the model. The reduction from 200 to 100 basking sharks during parameter selection was necessary for computational feasibility. Reducing the number of agents during these exploratory stages allowed a broader range of parameter combinations to be screened within the available computational budget, with the full 200-shark, 50-trial testing reserved for confirmatory runs of the final selected settings.
Small trials (5–10) were then run, and the RMSE, ME, and MAE were calculated until six settings emerged as producing the most realistic output (Suppl. material
Average number of shark aggregations for 50 trials of Test A and Test B.
| Submodel | Total Aggregations | Pseudo-Sighting Reports | |
|---|---|---|---|
| Test A (best fit for Food/Social submodel) | Food/Social | 4099 | 208 |
| Social | 568 | 170 | |
| Random | 629 | 172 | |
| Test B (best fit for Food submodel) | Food | 73791 | 322 |
| Social | 105 | 178 | |
| Random | 173 | 170 |
Comparison of the average number of aggregations per month with IBSG/IWDG data.
| Submodel | Inishowen | All of Ireland | ||||||
|---|---|---|---|---|---|---|---|---|
| ME | RMSE | MAE | ME | RMSE | MAE | |||
| Test A (best fit for Food/Social submodel) | Total Aggregations | Food/Social | 0.03 | 0.16 | 0.06 | 0.0 | 0.17 | 0.08 |
| Social | 0.27 | 0.39 | 0.28 | 0.25 | 0.36 | 0.26 | ||
| Random | 0.16 | 0.24 | 0.17 | 0.13 | 0.21 | 0.16 | ||
| Pseudo-Sighting Reports | Food/Social | 0.11 | 0.18 | 0.13 | 0.06 | 0.19 | 0.13 | |
| Social | 0.23 | 0.28 | 0.24 | 0.18 | 0.27 | 0.22 | ||
| Random | 0.27 | 0.31 | 0.27 | 0.22 | 0.29 | 0.25 | ||
| Test B (best fit for Food submodel) | Total Aggregations | Food | 0.46 | 0.51 | 0.47 | 0.43 | 0.49 | 0.45 |
| Social | 0.10 | 0.18 | 0.12 | 0.07 | 0.17 | 0.11 | ||
| Random | 0.14 | 0.24 | 0.16 | 0.11 | 0.22 | 0.15 | ||
| Pseudo-Sighting Reports | Food | 0.17 | 0.21 | 0.18 | 0.14 | 0.21 | 0.18 | |
| Social | 0.25 | 0.32 | 0.26 | 0.22 | 0.29 | 0.23 | ||
| Random | 0.30 | 0.37 | 0.31 | 0.28 | 0.35 | 0.29 | ||
Preliminary testing identified Test A as the best parameter setting for the Food/Social submodel, while Test B was identified as the best parameter setting for the Food submodel. Fifty trials were run under these parameter settings, with the output compared with the IBSG/IWDG sightings data.
Once the settings that produced the most realistic outputs for the Food/Social (Test A) and Food (Test B) submodels were identified, 50 trials were run for each setting (Table
Normalized results from Tests A and B were plotted as box plots and time-series graphs alongside normalized results from Inishowen and Ireland. The average size of shark aggregations per day (derived from Total Aggregations) was graphed in a scatter plot against the average amount of zooplankton in the same patches on a log scale.
Sensitivity and robustness analysis (SA/RA) testing (
Sensitivity analyses were conducted by altering the settings of Test A one by one, each by ±5%, ±35%, and ±70%, as described by
Robustness testing was conducted by removing, one at a time, the following five behavioral mechanisms from the Food/Social submodel: sharks did not check food availability in their patch; sharks did not seek other sharks; sharks did not “migrate” out of the model area when they had not eaten after a set number of days; sharks did not maintain a memory of high-zooplankton patches; and sharks did not seek patches of high zooplankton. The percent difference in the average number of aggregations and average maximum aggregation size was calculated between the robustness analysis and the results of Test A.
Exploratory testing narrowed the range of parameter settings that produced semi-realistic results for the Food and Food/Social submodels. No tests produced realistic results for either the Social or Random submodels. From the exploratory testing, it was determined that setting Sense-Distance to 10, Swim-Speed to 9, Friend_Min to 5, and Return-Season to 20 produced the most realistic results for both the Food and Food/Social submodels (Table
The number of Total Aggregations in Test A (best fit for the Food/Social submodel) ranged from 1,712–3,925 aggregations for the Food submodel but had a lower range for the Social (143–568 aggregations) and Random submodels (165–629 aggregations; Table
When comparing the normalized number of aggregations per month (averaged across all trials) from the model tests with the IBSG/IWDG data, the Food/Social submodel produced the lowest ME, RMSE, and MAE results (0.0–0.17) of any submodel in either Test A or Test B (Table
A comparison of the time series of the normalized number of aggregations per month (averaged across all trials) for both Test A and Test B shows that the Food/Social submodel in Test A produced the most realistic time-series output, although it was not a perfect match to the IBSG/IWDG data. However, the Food/Social submodel produced more episodic, irregular increases and decreases in the number of aggregations, while the Food submodel produced more rhythmic or seasonal cycling that was less comparable to the IBSG/IWDG data. This was exhibited by both the Total Aggregations and Pseudo-Sighting Reports. When considering the box plots, only the Food/Social submodel (Test A) produced an interquartile range that overlapped with data from both Ireland and Inishowen, although the normalized Total Aggregations displayed a more comparable range than the Pseudo-Sighting Reports (Fig.
Total number of aggregations throughout the entire model run (1982–2018) for Test A and Test B. Pseudo-Sighting Reports include any shark “sighted” during a random sample of 20 patches, including single sharks, while Total Aggregations count only groups of two or more sharks but count all aggregations in the model each day. See Table
Average number of aggregations per month for Test A. Comparison of the average number of aggregations per month throughout the model run (1982–2018), averaged across 50 trials for Test A (see Table
Average number of aggregations per month for Test B. Comparison of the average number of aggregations per month throughout the model run (1982–2018), averaged across 50 trials for Test B (see Table
Box plot of the average number of aggregations and average aggregation size per month for Test A. Box plots comparing the average number of aggregations and average size of aggregations per month throughout the model run (1982–2018), averaged across 50 trials for Test A (see Table
Box plot of average number of aggregations and average size per month for Test B. Box plots comparing the average number of aggregations and average size of aggregations per month throughout the model run (1982–2018), averaged across 50 trials for Test B (see Table
Average size of aggregations per month for Test A. Comparison of the average size of aggregations per month throughout the model run (1982–2018), averaged across 50 trials for Test A (see Table
Comparison of the average number of aggregations per month, averaged across 50 trials for Tests A and B (see Table
When comparing the normalized monthly aggregation size (averaged across all trials) across all tests, the Food/Social submodel (Test A) resulted in the lowest RMSE value (0.18; Table
| Submodel | Inishowen | All of Ireland | ||||||
|---|---|---|---|---|---|---|---|---|
| ME | RMSE | MAE | ME | RMSE | MAE | |||
| Test A (best fit for Food/Social submodel) | Total Aggregations | Food/Social | 0.10 | 0.18 | 0.13 | 0.08 | 0.18 | 0.12 |
| Social | 0.62 | 0.70 | 0.63 | 0.61 | 0.68 | 0.61 | ||
| Random | 0.58 | 0.67 | 0.58 | 0.56 | 0.65 | 0.57 | ||
| Pseudo-Sighting Reports | Food/Social | 0.07 | 0.16 | 0.11 | 0.05 | 0.15 | 0.10 | |
| Social | 0.38 | 0.43 | 0.39 | 0.36 | 0.41 | 0.37 | ||
| Random | 0.43 | 0.48 | 0.44 | 0.41 | 0.45 | 0.42 | ||
| Test B (best fit for Food submodel) | Total Aggregations | Food | 0.17 | 0.21 | 0.18 | 0.14 | 0.21 | 0.18 |
| Social | 0.25 | 0.32 | 0.26 | 0.22 | 0.29 | 0.23 | ||
| Random | 0.30 | 0.37 | 0.31 | 0.28 | 0.35 | 0.29 | ||
| Pseudo-Sighting Reports | Food | 0.16 | 0.27 | 0.19 | 0.14 | 0.26 | 0.18 | |
| Social | 0.36 | 0.43 | 0.37 | 0.34 | 0.40 | 0.35 | ||
| Random | 0.35 | 0.40 | 0.40 | 0.33 | 0.38 | 0.34 | ||
The Food/Social submodel in Test A displayed the strongest qualitative similarities to the timing of the IBSG/IWDG data, following a nonuniform but seasonal pattern of aggregation size through time. However, increases in the average aggregation size occurred in the model with a slight lag compared with the IBSG/IWDG data (Fig.
Average size of aggregations per month for Test B. Comparison of the average size of aggregations per month throughout the model run (1982–2018), averaged across 50 trials for Test B (see Table
Graphing the average daily aggregation size against the average amount of zooplankton in the same patches showed that zooplankton availability had no impact on aggregation size in the Social or Random submodels. Under the settings of Test A, the Food/Social submodel showed a slight trend, with the majority of aggregations being smaller across a wider range of zooplankton availability (Fig.
Comparison of the average size of aggregations per month, averaged across 50 trials for Tests A and B (see Table
The results of 50 trials were averaged together. The scatter plot depicts the average aggregation size per day (taken from Total Aggregations) compared with the average amount of zooplankton in the patches where the aggregations were observed. Zooplankton abundance was graphed on a logarithmic scale.
The model parameter with the highest impact on the number of aggregations in the Food/Social submodel was Cal_%, with Return-Season having the strongest impact on the size of the aggregations. Return-Season and Threshold_ZP also had a strong impact on the number of aggregations, with No_Eat_Min having an impact on the size of aggregations (Table
| Average number of aggregations | ||||||
|---|---|---|---|---|---|---|
| Parameter | –70% | –35% | –5% | +5% | +35% | 70% |
| Cal_% | –7.36 | –7.36 | –2.10 | –3.09 | –1.34 | –1.00 |
| Friend_Min | –0.74 | –0.74 | –0.42 | –0.17 | –0.05 | –0.07 |
| No_Eat_Min | 0.30 | 0.30 | 1.03 | –0.49 | 1.09 | 0.50 |
| Other_ZP_% | –0.41 | –0.41 | 0.19 | –0.23 | –0.03 | –0.01 |
| Return-Season | –2.69 | –2.69 | –1.99 | –6.84 | –1.23 | –0.71 |
| Sense-Distance | 0.79 | 0.79 | –0.13 | –1.26 | –0.9 | –0.70 |
| Swim-Speed | 1.14 | 1.14 | 1.65 | 0.77 | 0.22 | 0.15 |
| Threshold_ZP | –6.18 | –6.18 | –1.30 | –3.50 | –1.36 | –0.94 |
| Average maximum aggregation size | ||||||
| Cal_% | –2.46 | –1.73 | –0.75 | 0.17 | –0.26 | –0.74 |
| Friend_Min | 0.01 | -0.18 | 0.07 | –0.06 | –0.09 | –0.14 |
| No_Eat_Min | 0.46 | -0.82 | –1.89 | –1.75 | 1.97 | 0.89 |
| Other_ZP_% | 0.00 | -0.23 | –0.96 | 1.59 | 0.01 | –0.05 |
| Return-Season | –0.47 | 0.13 | –5.95 | –5.50 | –1.17 | –0.58 |
| Sense-Distance | 1.20 | 1.28 | 0.42 | –0.93 | –0.15 | –0.13 |
| Swim-Speed | 1.27 | 1.42 | 1.48 | 0.93 | 0.42 | 0.16 |
| Threshold_ZP | –2.23 | –1.77 | 1.28 | 1.52 | 0.2 | –0.27 |
Removing No_Eat_Min had the most significant impact on both the number and size of aggregations (Table
Robustness analysis. Robustness testing was conducted on the Food/Social submodel with the parameter settings of Test A. The listed behaviors were removed one at a time, and 10 trials were run for each behavior that was removed. The results of 10 trials were averaged. The percent difference in the average number of aggregations and average maximum aggregation size was calculated between the robustness analysis and the results of Test A. The 95% confidence intervals were calculated using a t-distribution (df = 9) across the 10 trials.
This model demonstrates the viability of IBMs as a mechanism for modeling basking shark aggregations. The results of the different submodels and the robustness analysis make clear that the behavior of the simulated sharks has a clear impact on the model output, demonstrating that aggregations were not a result of random chance but of shark (agent) behavior.
While the Social and Random submodels had higher consistency between trials, this is because both models rarely produced aggregations of more than two sharks. In the Social submodel, basking sharks aggregate based on conspecific cues alone, without assessing food availability in a patch. Because there is no shared environmental cue drawing multiple individuals toward the same location, sharks continuously reassess and move toward whichever nearby individuals satisfy their social criteria, but with no environmental factors causing multiple sharks to converge on or near the same area. The fact that this submodel rarely produces aggregations larger than two individuals suggests that social attraction in isolation is insufficient to generate the larger aggregations observed in the IBSG/IWDG data.
In contrast, the more ecologically complete model (Food/Social submodel) resulted in lower RMSE, ME, and MAE results and more realistic time series compared with the Social and Random submodels. The submodel that produced results that were most statistically and qualitatively similar to the IBSG/IWDG data was the submodel that included both food availability and social interactions as drivers of basking shark behavior, resulting in the emergence of aggregations not unlike those observed in the wild. This is consistent with research indicating that basking shark aggregations are the result of a combination of food availability and social interactions (
The removal of certain behaviors in the robustness analysis had a significant impact on the model outputs. Removing the behavior of seeking other sharks increased both the number and size of the aggregations, as demonstrated by the robustness tests and the results of the Food submodel in Test B. Notably, the Food submodel resulted in trends that were a weaker fit to the IBSG/IWDG data, indicating that food availability is insufficient to explain these aggregations.
Migration from the surface waters of the model area clearly impacted the results of the model, as demonstrated by the results of the robustness tests. Removing this behavior greatly inflated the number of aggregations, as 200 sharks remained in the model area for the entirety of each modeled year, an unrealistic behavior for sharks in this area (
When basking sharks did not check a patch for sufficient food to support the sharks in the area, the model resulted in fewer aggregations and smaller aggregation sizes, as sharks did not receive a simulated environmental cue to leave a patch. Similarly, when sharks did not seek out zooplankton, their movements were largely random, resulting in fewer and smaller aggregations. A lack of memory means that the shark is only responding directly to environmental cues and not making any other decisions, something that is unlikely in such a long-lived species that migrates long distances (
The impact of food availability on aggregations was clearly demonstrated in the model, in agreement with the literature (
Research on the threshold level of zooplankton required for basking sharks to both move to and remain in an area is still limited (
The preliminary tests that had a lower percentage of patches with zooplankton (Cal_% and Other_ZP_% < 20%) resulted in aggregation trends that were more likely to match the IBSG/IWDG sightings data. This indicates that the Inishowen Peninsula may be an area of high zooplankton patchiness and that zooplankton patchiness increases the likelihood of these aggregations. This is further demonstrated by the results of Test B, where only 10% of patches contained zooplankton. Test B had a notably higher number of aggregations in the Food submodel than were documented in the Food/Social submodel (Test A). This is indicative of food availability and the location of food being drivers of aggregations within the model. This may be a reason why the Food/Social submodel performed better than the Food submodel, while in tests with a higher percentage of zooplankton patches, those two submodels were virtually identical (Suppl. material
Utilizing data from field studies, a rough estimate of the average daily (straight-line) distance traveled by a basking shark was calculated to be approximately 9 km (
The accuracy of the Sense-Distance parameter is difficult to assess, as the mechanism by which basking sharks sense zooplankton or other sharks is still poorly understood, and the Sense-Distance is likely impacted by the distance a shark can move within a 24 h period. While sharks can sense sound over thousands of meters, they may only be able to sense smells within hundreds of meters (
This model has several limitations. Most notably, the lack of real-world localized zooplankton data makes it difficult, if not impossible, to assess how realistic the surface environment in the model is for the sharks. Therefore, it is virtually impossible to make a realistic comparison of the location of basking sharks within the model with the IBSG/IWDG sightings data. However, the current model demonstrates that the inclusion of zooplankton is a workable and important component for improving realism when modeling basking sharks. Next steps should involve improved zooplankton data input. Further research could combine satellite phytoplankton data, which can provide fine-scale density and distribution of chlorophyll, with CPR data that provide insight into the abundance and species of zooplankton to simulate zooplankton location and patchiness.
CPR sampling alone does not reflect localized patchiness but samples across transects, so field research intended to assess zooplankton should be conducted in the modeled area to ensure accuracy. New and emerging technology (such as digital holographic cameras and the use of satellite color detection) will also allow future research to obtain a clearer picture of the in situ preyscape for basking sharks (
The memory mechanism is constrained by the current representation of zooplankton in the model. In the model’s current form, zooplankton are randomly distributed each day rather than following a more realistic temporal or spatial distribution. Therefore, the memory mechanism is not able to record a plausible seasonal pattern or long-term trends, thereby limiting its impact on shark foraging behavior. Given this constraint, the ecological significance of memory-based foraging remains largely hypothetical, and the relatively small, though nonzero, effect of removing memory in the robustness analysis should be interpreted with this limitation in mind. Improved zooplankton simulations will allow more thorough testing of the impact of this mechanism on shark behavior.
This research also shows the limitations of publicly reported sightings, as the Pseudo-Sighting Reports in all model versions missed a sizable number of aggregations, even with random sampling of 20 patches per day (a consistent effort not seen in the IBSG/IWDG data). The model data are comparably more systematic than the IBSG/IWDG sighting reports, which are not systematic surveys but rely on individuals to recognize a basking shark and know to report it. Research in 2018 in Inishowen demonstrated that a large majority of people (both tourists and residents) were unaware that they could report a sighting (
This model is also limited by its focus on surfacing sharks. This is because the sightings data face the same limitation. However, advances in technology, including increasingly complex and data-heavy tags, will result in a deeper understanding of basking sharks’ subsurface behavior. This IBM could then be expanded to incorporate subsurface behavior, subsurface preyscapes, and bathymetric data, allowing for more holistic ecological modeling. In the future, this model can be duplicated in other areas where basking sharks are sighted, with moderate changes to the underlying map and zooplankton data.
As data on shark populations improve, the model can be improved by estimating death and birth rates, the age or sex of the sharks, and the number of sharks that are likely to exhibit site fidelity in the area. While these data are largely unavailable at the moment, research is currently underway to expand the knowledge base of sharks in Ireland.
Understanding how food drives both long-term migratory patterns and localized shark behavior is vital for conservation in the face of climate change, and individual-based modeling is a mechanism by which theories of behavior can be tested and, with improved environmental data, predictive models can be generated. Long-term sea surface temperature (SST) increases have been observed in the North Atlantic in previous decades, and multiple studies have found that this increase in SST has led to a northward shift in zooplankton communities, with an increase in the range of warm-water zooplankton and a decrease in the range of cold-water zooplankton, including C. finmarchicus (
Even with limited environmental data, this IBM was able to reproduce long-term trends found in publicly reported sightings. This model allows for flexibility in both environmental modeling and behavioral attributes, allowing for increasing complexity as new data become available, as well as for the testing of different hypotheses of behavior. Environmental research is still required to bolster the underlying environment, thereby creating even more realistic behaviors. The ability of IBMs to account for individualized shark behavior may become more necessary in light of current research focused on individualized social behavior of basking sharks (A. McInturf, personal communication, 8 August 2023), as well as other research demonstrating unique social groupings in other sharks (
The authors thank the Irish Basking Shark Group and Irish Whale and Dolphin Group for sharing the long-term sightings datasets. The authors also thank the Jackson Laboratory and George Mason University for computing resources.
The authors have declared that no competing interests exist.
No ethical statement was reported.
The authors accept full responsibility for the content of the manuscript, including the disclosure of any use of AI.
No AI tools were used in the preparation of this manuscript.
No funding was reported.
Conceptualization: CS, EPB, CG, EP. Data curation: ACG, CG. Formal analysis: CG. Investigation: DSR. Methodology: CG, ACG, DSR. Software: ACG, DSR. Supervision: EP, DSR, EPB, CS. Validation: CG. Visualization: CG. Writing - original draft: CG. Writing - review and editing: CG, CS, ACG, DSR, EP, EPB.
C. Gray https://orcid.org/0000-0001-9441-7242
D.S. Rothman https://orcid.org/0000-0003-1479-2849
E. Peters-Burton https://orcid.org/0000-0003-1233-9119
E.C.M Parsons https://orcid.org/0000-0002-0464-1046
A. Garretson https://orcid.org/0000-0002-7260-0131
All of the data that support the findings of this study are available in the main text or Supplementary Information.
Basking Shark IBM
Data type: zip
Explanation note: Contains IBM (netlogo file) and necessary input data to run the model.
Model documentation
Data type: docx
Supplemental model results
Data type: docx